JP2006126859A - Speech processing device and method - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a speech processing device and a speech processing method with which distortion in speech is made less and noise is sufficiently eliminated. <P>SOLUTION: A voice/nonvoice discriminating section 106 judges that a section of the voice spectrum is a voice section containing a voice component if the difference between the voice spectrum signal and the value of a noise base is a prescribed threshold or more and otherwise judges that the section is a nonvoice section containing no voice components and containing only noise. A comb filter generating section 107 generates a comb filter for enhancing the voice pitch according to whether or not a voice component is contained in each frequency bin. A damping coefficient calculating section 108 multiplies the comb filter by a damping coefficient, based on a frequency characteristic, determines the damping coefficient of the input signal for each frequency bin, and outputs the damping coefficient of each frequency bin to a multiplying section 109. The multiplying section 109 multiplies the voice spectrum by the damping coefficient for each frequency bin unit. A frequency synthesizing section 110 combines the spectra of the frequency bin units determined by the multiplication to synthesize a voice spectrum continuous in a frequency range in units of a prescribed processing time. <P>COPYRIGHT: (C)2006,JPO&NCIPI

Description

  The present invention relates to a speech processing apparatus and speech processing method for suppressing noise, and more particularly to a speech processing apparatus and speech processing method in a communication system.

  With conventional speech coding technology, it is possible to talk with high-quality speech for speech without noise, but for speech with noise, harsh noise unique to digital speech occurs, resulting in There was a problem of deterioration.

  As a speech enhancement technique for suppressing such noise, there are a spectral subtraction method and a comb filter method.

  Spectral subtraction method focuses on noise information, estimates the nature of the noise in the silent period, and subtracts the short-time power spectrum of the noise from the short-time power spectrum of the speech signal containing noise, or multiplies the attenuation coefficient. This is a method of suppressing noise by estimating the power spectrum of an audio signal. Spectral subtraction methods are described in, for example, literature (S. Boll, Suppression of acoustic noise in speech using spectral subtraction, IEEE Trans. Acoustics, Speech, and Signal Processing, vol. ASSP-27, pp. 113-120, 1979), literature. (RJMcAulay, MLMalpass, Speech enhancement using a soft-decision noise suppression filter, IEEE.Trans.Acoustics, Speech, and Signal Processing, vol.ASSP-28, pp.137-145. 1980), Patent No. 2714656 And Japanese Patent Application No. 9-518820.

  On the other hand, the comb filter method focuses on voice information and performs noise attenuation by applying a comb filter to the pitch of the voice spectrum. For example, the literature (JSLim etc., Evaluation of an adaptive comb filtering method for enhancing speech degraded by white noise addition, IEEE Trans.Acoustics, Speech, and Signal Processing, vol.ASSP26, pp.354 -358,1978).

  A comb filter is a filter that attenuates a signal input in a frequency domain unit at a predetermined ratio or outputs a signal without attenuation, and has a comb-like attenuation characteristic. When the comb filter method is realized by digital data processing, it is possible to suppress noise by creating attenuation characteristic data for each frequency domain and multiplying a voice spectrum for each frequency.

  FIG. 28 is a diagram illustrating an example of a speech processing apparatus using a conventional comb filter method. In FIG. 28, the switcher 11 outputs the input signal as it is when the input signal includes a speech component having no quasi-periodicity (for example, consonant), and the speech signal having the quasi-periodicity is included in the input signal. If so, the input signal is output to the comb filter 12. The comb filter 12 attenuates the noise part in the frequency domain with respect to the input signal with an attenuation characteristic based on the pitch period information, and outputs the attenuated signal.

  FIG. 29 is a diagram illustrating attenuation characteristics of the comb filter. The vertical axis represents the signal attenuation characteristics, and the horizontal axis represents the frequency. In FIG. 29, the comb filter has a region where the signal is attenuated and a region where the signal is not attenuated for each frequency region.

  In the comb filter method, by applying a comb filter to the input signal, the frequency region where the audio component exists in the input signal is not attenuated, and the noise is suppressed by attenuating the frequency region where the audio component does not exist. To emphasize the voice.

  However, such a conventional voice processing method has the following problems to be solved. First, the SS method shown in Document 1 focuses only on noise information, regards short-time noise characteristics as steady, does not distinguish between speech and noise, and is uniformly noise-based (estimated noise spectral characteristics). It is a method of subtracting. Audio information (for example, audio pitch) is not used. Actually, since the noise characteristics are not steady, the residual noise after subtraction, especially the residual noise between pitch harmonics, may cause unnatural distortion noise called so-called `` musical noise '' depending on the processing method. Conceivable.

  As an improvement method, a method of attenuating noise by multiplying an attenuation coefficient based on the voice power to noise power ratio (SNR), for example, those shown in Japanese Patent No. 2714656 and Japanese Patent Application No. 9-518820 are proposed. It was. Since different attenuation coefficients are used by distinguishing between a high voice band (SNR is high) and a high noise band (SNR is low), musical noise is suppressed and sound quality is improved. However, although the method shown in Japanese Patent No. 2714656 and Japanese Patent Application No. 9-518820 uses a part of audio information (SNR), the number of frequency channels (16 channels) to be processed is not sufficient. It is difficult to separate and extract pitch harmonic information from noise, and since attenuation coefficients are used in both voice and noise bands, the attenuation coefficient cannot be increased as a result of mutual influence. That is, when the attenuation coefficient is increased, there is a possibility that voice distortion may occur due to an SNR estimation error. As a result, noise attenuation is insufficient.

  Further, in the conventional comb filter method, if there is an estimation error in the pitch that is the fundamental frequency, the amount of error increases in the harmonic, and the possibility that the original harmonic component deviates from the pass band becomes greater. In addition, since it is necessary to distinguish between speech having quasi-periodicity and speech that is not so, there is a problem in feasibility.

  The present invention has been made in view of this point, and an object of the present invention is to provide an audio processing apparatus and an audio processing method that can reduce noise and sufficiently eliminate noise.

  The speech processing apparatus according to the present invention includes a frequency dividing unit that divides a speech spectrum of an input signal in units of a predetermined frequency region, and a speech component is included in the divided speech spectrum that is a speech spectrum frequency-divided by the frequency dividing unit. A voice / non-speech identification means for identifying whether or not the speech / non-speech identification means is used, and a frequency region including a speech component is defined as a pass band based on the identification result of the speech / non-speech identification unit, and a frequency region including no speech component is blocked A comb filter generating means for generating a comb filter as a region, and an attenuation coefficient for each frequency domain is set by multiplying the comb filter by an attenuation coefficient based on a frequency characteristic, and corresponding to each of the divided speech spectrums Noise suppression means for suppressing a noise component of the divided speech spectrum by multiplying by a frequency domain attenuation coefficient; and the noise component A configuration that includes a frequency synthesis means for synthesizing the speech spectrum the oppressed divided speech spectrum continuous in the frequency region.

  The audio processing program of the present invention includes an audio component in a frequency division procedure for dividing an audio spectrum of an input signal in units of a predetermined frequency domain, and a divided audio spectrum that is a frequency spectrum divided in the frequency division procedure. The voice / non-speech identification procedure for identifying whether or not the speech / non-speech is identified, and the frequency region including the speech component based on the identification result of the speech / non-speech identification procedure is set as the pass band, and the frequency region not including the speech component is blocked. A comb filter generation procedure for generating a comb filter as a region, and an attenuation coefficient for each frequency domain is set by multiplying the comb filter by an attenuation coefficient based on a frequency characteristic, and the corresponding corresponding to each divided speech spectrum A noise suppression procedure for suppressing a noise component of the divided speech spectrum by multiplying by a frequency domain attenuation coefficient; Including a frequency synthesis procedure for synthesizing the speech spectrum component is continuously suppressed frequency domain divided speech spectra were, a.

  The audio processing method of the present invention includes a frequency dividing step of dividing an audio spectrum of an input signal in units of a predetermined frequency region, and whether or not an audio component is included in the divided audio spectrum that is the frequency-divided audio spectrum Voice / non-speech identification step, a pitch harmonic structure generation step for generating a pitch harmonic structure for emphasizing a frequency region including a voice component, and an attenuation coefficient based on frequency characteristics are multiplied by the pitch harmonic structure. An attenuation coefficient setting step for setting an attenuation coefficient for each frequency region, and a noise suppression step for suppressing a noise component of the divided speech spectrum by multiplying each of the divided speech spectra by an attenuation coefficient of the corresponding frequency region And a frequency synthesizer for synthesizing the divided speech spectrum with the noise component suppressed into a continuous speech spectrum in the frequency domain. Adopt a method that includes the, the.

  According to the present invention, a speech spectrum is identified in a frequency domain unit as a region with a speech component and a region without a speech component, and noise is suppressed based on a highly accurate pitch period obtained from this identification information. Distortion and noise can be sufficiently removed.

  The gist of the present invention is a comb that distinguishes a speech spectrum into a region with a speech component and a region without a speech component in a frequency domain unit and emphasizes only the speech information based on a highly accurate pitch period obtained from this identification information. The noise is suppressed by generating a filter in the frequency domain.

(Embodiment 1)
FIG. 1 is a block diagram showing the configuration of the speech processing apparatus according to Embodiment 1 of the present invention. In FIG. 1, the speech processing apparatus includes a time division unit 101, a windowing unit 102, an FFT unit 103, a frequency division unit 104, a noise base estimation unit 105, a speech non-speech identification unit 106, and a comb filter generation. Unit 107, attenuation coefficient calculation unit 108, multiplication unit 109, frequency synthesis unit 110, and IFFT unit 111.

  The time division unit 101 configures a frame divided in predetermined time units from the input audio signal, and outputs the frame to the windowing unit 102. The windowing unit 102 performs window processing using a Hanning window on the frame output from the time division unit 101 and outputs the result to the FFT unit 103. The FFT unit 103 performs FFT (Fast Fourier Transform) on the audio signal output from the windowing unit 102, and outputs the audio spectrum signal to the frequency division unit 104.

  The frequency division unit 104 divides the speech spectrum output from the FFT unit 103 into frequency components in a predetermined frequency domain unit, and the speech spectrum is divided into a noise base estimation unit 105, a speech non-speech identification unit 106, and the like for each frequency component. To the multiplication unit 109. The frequency component indicates a voice spectrum divided in a predetermined frequency unit.

  When the determination result that the speech component is included in the frame is output from the speech non-speech identification unit 106, the noise base estimation unit 105 outputs the noise base estimated in the past to the speech non-speech discrimination unit 106. In addition, when the determination result that the speech component is not included in the frame is output from the speech non-speech identification unit 106, the noise base estimation unit 105 is a short time for each frequency component of the speech spectrum output from the frequency division unit 104. A moving average value representing an average amount of power spectrum and spectrum change is calculated, a moving average value calculated in the past and a weighted average value of the power spectrum are calculated, and a new moving average value is calculated.

Specifically, the noise base in each frequency component is estimated using Equation (1) and output to the speech non-speech identification unit 106.

Here, n is a number that identifies a frame to be processed, k is a number that identifies a frequency component, and τ is a delay time. S ² _f (n, k) represents the power spectrum of the input audio signal, P _base (n, k) represents a noise-based moving average value, and α (k) represents a moving average coefficient.

  The speech non-speech identification unit 106, when the difference between the speech spectrum signal output from the frequency division unit 104 and the noise base value output from the noise base estimation unit 105 is greater than or equal to a predetermined threshold value, It is determined to be a part, and otherwise, it is determined to be a silent part of only noise that does not include a voice component. Then, the speech non-speech identification unit 106 outputs the determination result to the noise base estimation unit 105 and the comb filter generation unit 107.

  Comb filter generation section 107 generates a comb filter that emphasizes pitch harmonics based on the presence or absence of a voice component in each frequency component, and outputs this comb filter to attenuation coefficient calculation section 108. Specifically, the comb filter generation unit 107 turns on the frequency component of the sound part of the comb filter and turns off the frequency component of the silent part.

  The attenuation coefficient calculation unit 108 multiplies the comb filter generated by the comb filter generation unit 107 by an attenuation coefficient based on frequency characteristics, and sets the attenuation coefficient of the input signal for each frequency component. Are output to the multiplier 109.

For example, the attenuation coefficient gain (k) can be calculated from the following equation (2) and multiplied by the input signal.

  Here, gc is a constant, k is a variable for specifying a bin, and HB is an FFT transform length, that is, the number of data to be subjected to fast Fourier transform.

  Multiplier 109 multiplies the audio spectrum output from frequency divider 104 by the attenuation coefficient output from attenuation coefficient calculator 108 in units of frequency components. Then, the spectrum obtained as a result of multiplication is output to frequency synthesizer 110.

  The frequency synthesis unit 110 synthesizes the spectrum of the frequency component unit output from the multiplication unit 109 into a speech spectrum continuous in the frequency domain in a predetermined processing time unit and outputs the synthesized speech spectrum to the IFFT unit 111. The IFFT unit 111 performs an IFFT (Inverse Fast Fourier Transform) on the audio spectrum output from the frequency synthesis unit 110 and outputs a signal converted into an audio signal.

  Next, the operation of the speech processing apparatus having the above configuration will be described using the flowchart shown in FIG. In FIG. 2, in step (hereinafter referred to as “ST”) 201, preprocessing is performed on the input signal. In this case, the preprocessing is to perform a windowing process by constructing a frame of a predetermined time unit from the input signal, and to perform a fast Fourier transform on the speech spectrum.

  In ST202, frequency division section 104 divides the audio spectrum into frequency components. In ST203, the noise base estimation unit 105 determines whether α (k) = 0, that is, whether to stop the noise base update. If α (k) = 0, the process proceeds to ST205. If α (k) = 0 is not satisfied, the process proceeds to ST204.

In ST204, the noise base estimation unit 105 updates the noise base from a speech spectrum that does not include a speech component, and then proceeds to ST205. In ST205, the audio non-speech identifying section ^{_{106, S f 2 (n, k}} )> Q up · P base (n, k) is whether, i.e. the power of the speech spectrum is multiplied by a predetermined threshold noise base If S _f ² (n, k)> Q _up · P _base (n, k), the process proceeds to ST206 and S _f ² (n, k)> Q _up · If not P _base (n, k), the process proceeds to ST208.

In ST206, the speech non-speech identification unit 106 sets α (k) = 0 indicating that the noise base update is stopped. In ST207, the comb filter generation unit 107 sets SP_SWITCH (k) = ON indicating that the voice spectrum is output without being attenuated, and the process proceeds to ST211. In ST208, the speech non-speech identification unit 106 determines whether or not S _f ² (n, k) <Q _down · P _base (n, k), that is, the power of the speech spectrum multiplies the noise base by a predetermined threshold value. If S _f ² (n, k) <Q _down · P _base (n, k), the process proceeds to ST209 and S _f ² (n, k) <Q _down · If not P _base (n, k), the process proceeds to ST211.

  In ST209, the speech non-speech identification unit 106 sets α (k) = SLOW indicating noise base update. Here, SLOW is a predetermined constant. In ST210, SP_SWITCH (k) = OFF indicating that the comb filter generation unit 107 attenuates and outputs the voice spectrum is set, and the process proceeds to ST211.

  In ST211, attenuation coefficient calculation section 108 determines whether or not the audio spectrum is not attenuated, that is, whether SP_SWITCH (k) = ON. When SP_SWITCH (k) = ON in ST211, in ST212, the attenuation coefficient calculation unit 108 sets the attenuation coefficient to 1, and the process proceeds to ST214. If ST_SWITCH (k) is not ON in ST211, in ST213, the attenuation coefficient calculation unit 108 calculates and sets an attenuation coefficient corresponding to the frequency, and proceeds to ST214.

  In ST214, multiplication section 109 multiplies the audio spectrum output from frequency division section 104 by the attenuation coefficient output from attenuation coefficient calculation section 108 in units of frequency components. In ST215, the frequency synthesizer 110 synthesizes the frequency component unit spectrum output from the multiplier 109 into a continuous speech spectrum in the frequency domain in a predetermined processing time unit. In ST216, IFFT section 111 performs IFFT on the speech spectrum output from frequency synthesis section 110 and outputs a signal in which noise is suppressed.

  Next, a comb filter used in the speech processing apparatus according to the present embodiment will be described. FIG. 3 is a diagram illustrating an example of a comb filter created by the speech processing apparatus according to the present embodiment. In FIG. 3, the vertical axis indicates the power of the spectrum and the attenuation of the filter, and the horizontal axis indicates the frequency.

  The comb filter has an attenuation characteristic indicated by S1, and the attenuation characteristic is set for each frequency component. The comb filter generation unit 107 attenuates a frequency domain signal that does not include an audio component, and creates an attenuation characteristic comb filter that does not attenuate a frequency domain signal that includes an audio signal.

  The speech spectrum S2 including the noise component is subjected to a comb filter having an attenuation characteristic of S1, so that the frequency domain signal including the noise component is attenuated and the power is reduced, and the portion including the audio signal is not attenuated and the power is increased. It does not change. The obtained speech spectrum has a spectrum shape in which the frequency region of the noise component is lower and the peak is not lost and is emphasized, and the speech spectrum S3 in which the noise that does not lose the pitch harmonic information is suppressed is output.

  Thus, according to the speech processing apparatus according to Embodiment 1 of the present invention, speech non-speech in the spectrum signal is discriminated in units of frequency components, and frequency characteristics are attenuated based on the discrimination results in units of frequency components. Thus, accurate pitch information can be obtained, so that speech enhancement with little speech distortion can be performed even if noise suppression is performed with a large attenuation.

  Further, by providing two threshold values in voice identification, it is possible to discriminate voice non-voice with high accuracy.

  Note that the attenuation coefficient calculation unit 108 calculates the attenuation coefficient according to the frequency characteristics of the noise, so that speech enhancement can be performed without impairing the consonant at a high frequency.

  In addition, when the input signal is attenuated in binary for each frequency component and determined to be speech, attenuation is not performed, and when it is determined to be noise, attenuation can be performed. In this case, even if strong noise suppression is performed, a certain frequency component of the voice is not attenuated, so that voice enhancement with less voice distortion can be performed.

(Embodiment 2)
FIG. 4 is a block diagram of an example of the configuration of the speech processing apparatus according to the second embodiment. However, the same components as those in FIG. 1 are denoted by the same reference numerals as those in FIG.

  The voice processing apparatus of FIG. 4 includes a noise section discriminating unit 401 and a noise base tracking unit 402, performs voice / non-speech discrimination of a signal in units of frames, detects a sudden change in noise level, and promptly 1 is different from FIG. 1 in that the noise base is estimated and updated.

  In FIG. 4, the FFT unit 103 performs FFT (Fast Fourier Transform) on the audio signal output from the windowing unit 102, and outputs the audio spectrum to the frequency division unit 104 and the noise section determination unit 401.

  The noise section determination unit 401 calculates the signal power and moving average value in units of frames from the sound spectrum output from the FFT unit 103, and determines whether or not the frame includes sound from the rate of change of the power of the input signal. .

Specifically, the noise section determination unit 401 calculates the rate of change of the power of the input signal using the following formulas (3) and (4).

Here, P (n) is the signal power of one frame, S ² _f (n, k) is the input signal power spectrum, Ratio is the signal power ratio of the previously processed frame to the processed frame, τ is a delay time.

  The noise section determination unit 401 determines that the input signal is an audio signal when the ratio exceeds a preset threshold value for a certain period of time, and determines that the input signal is a noise section when the ratio does not exceed the threshold value continuously.

  When it is determined that the noise base tracking unit 402 has moved from the speech section to the noise section, the noise base tracking unit 402 increases the degree of influence of the noise base estimation from the processing frame in the noise base update while performing the processing of a predetermined number of frames.

  Specifically, α (k) = FAST and (0 <SLOW <FAST <1) are set in equation (1). The larger the value of α (k), the more easily the moving average value is affected by the input audio signal, and it is possible to cope with a sudden change in noise base.

  When the determination result that the speech component is not included in the frame is output from the speech non-speech discrimination unit 106 or the noise base tracking unit 402, the noise base estimation unit 105 outputs the frequency component of the speech spectrum output from the frequency division unit 104. A moving average value representing the average amount of the short-time power spectrum and the spectrum change for each time is calculated, and a noise base in each frequency component is estimated from these values and output to the speech non-speech identification unit 106.

  As described above, according to the speech processing apparatus according to Embodiment 2 of the present invention, by updating the noise base by largely reflecting the value of the noise spectrum estimated from the input signal, the noise level is rapidly increased. It is possible to perform a noise-based update corresponding to a change, and to perform speech enhancement with little speech distortion.

(Embodiment 3)
FIG. 5 is a block diagram of an example of the configuration of the speech processing apparatus according to the third embodiment. However, the same components as those in FIG. 1 are denoted by the same reference numerals as those in FIG.

  The speech processing apparatus of FIG. 5 includes a musical noise suppression unit 501 and a comb filter correction unit 502. When the frame includes sudden noise, the generated comb filter is corrected and the musical due to the sudden noise is generated. The point which suppresses generation | occurrence | production of noise differs from FIG.

  In FIG. 5, the comb filter generation unit 107 generates a comb filter that emphasizes pitch harmonics based on the presence or absence of a voice component in each frequency component, and outputs the comb filter to the musical noise suppression unit 501 and the comb filter correction unit 502.

  The musical noise suppression unit 501 is on in the state of each frequency component of the comb filter output from the comb filter generation unit 107, that is, when the number of states to be output without attenuation is equal to or less than a certain threshold value, It is determined that sudden noise is included in the frame, and the determination result is output to the comb filter correction unit 502.

For example, if the number of frequency components turned on by the comb filter is calculated using the following equation (5), and COMB_SUM (n) is smaller than a certain threshold value (for example, 10), musical noise is generated. to decide.

  The comb filter correction unit 502 corrects the comb filter to prevent the generation of musical noise based on the comb filter generation result output from the comb filter generation unit 107 in which sudden noise is included in the frame from the musical noise suppression unit 501. The comb filter is output to the attenuation coefficient calculation unit 108.

  Specifically, the state of all frequency components of the comb filter is set to OFF, that is, the signal is attenuated and output, and the comb filter is output to the attenuation coefficient calculation unit 108.

  The attenuation coefficient calculation unit 108 multiplies the comb filter output from the comb filter correction unit 502 by an attenuation coefficient based on frequency characteristics, sets the attenuation coefficient of the input signal for each frequency component, and sets each frequency component. The attenuation coefficient is output to the multiplier 109.

  As described above, according to the audio processing device according to Embodiment 3 of the present invention, by determining the occurrence of musical noise from the generation result of the comb filter, it is possible to prevent noise from being erroneously determined as an audio signal and It is possible to perform voice enhancement with little.

  Note that Embodiment 3 can be combined with Embodiment 2. That is, the effects of the second embodiment can be obtained by adding the noise section determination unit 401 and the noise base tracking unit 402 to the speech processing apparatus of FIG.

(Embodiment 4)
FIG. 6 is a block diagram of an example of the configuration of the speech processing apparatus according to the fourth embodiment. However, the same components as those in FIG. 1 are denoted by the same reference numerals as those in FIG. The speech processing apparatus in FIG. 6 includes an average value calculation unit 601 and is different from that in FIG. 1 in that an average value of speech spectrum power is obtained in units of frequency components.

  In FIG. 6, the frequency division unit 104 divides the audio spectrum output from the FFT unit 103 into frequency components indicating the audio spectrum divided by a predetermined frequency unit, and converts the audio spectrum into audio non-audio for each frequency component. The data is output to the identification unit 106, the multiplication unit 109, and the average value calculation unit 601.

  The average value calculation unit 601 takes the average value of the power of the voice spectrum output from the frequency division unit 104 with the frequency components in the vicinity and the average value with the previously processed frames, and uses the obtained average value as the noise base The data are output to the estimation unit 105 and the voice non-voice identification unit 106.

Specifically, the average value of the speech spectrum is calculated using the following formula (6).

  Here, k1 and k2 indicate frequency components, and k1 <k <k2. n1 indicates a number indicating a frame that has been processed in the past, and n indicates a number indicating a frame that has been processed.

  When the determination result that the speech component is not included in the frame is output from the speech non-speech identification unit 106, the noise base estimation unit 105 outputs, for each frequency component of the average value of the speech spectrum output from the average value calculation unit 601. A moving average value representing an average amount of short-time power spectrum and spectrum change is calculated, and a noise base in each frequency component is estimated and output to the speech non-speech identification unit 106.

  When the difference between the average value of the speech spectrum signal output from the average value calculation unit 601 and the noise base value output from the noise base estimation unit 105 is equal to or greater than a predetermined threshold, the speech non-speech identification unit 106 If the difference is smaller than a predetermined threshold value, it is determined that the sound is a silent part including only noise, and the determination result is determined as a noise base estimation unit 105 and a comb filter generation unit 107. Output to.

  As described above, according to the speech processing apparatus according to the fourth embodiment of the present invention, the average power value of the speech spectrum in each frequency component or the average power value of the frame processed in the past and the frame processed is obtained. Thus, the influence of the sudden noise component is reduced, and a more accurate comb filter can be configured.

  Note that Embodiment 4 can be combined with Embodiment 2 or Embodiment 3. That is, the effects of the second embodiment can be obtained by adding the noise section determination unit 401 and the noise base tracking unit 402 to the speech processing device of FIG. 5, and the musical noise suppression unit 501 and the speech processing device of FIG. The effect of the third embodiment can be obtained by adding the comb filter correction unit 502.

(Embodiment 5)
FIG. 7 is a block diagram of an example of the configuration of the speech processing apparatus according to the fifth embodiment. However, the same components as those in FIG. 1 are denoted by the same reference numerals as those in FIG.

  The speech processing apparatus in FIG. 7 includes a section determination unit 701 and a comb filter reset unit 702, and differs from FIG. 1 in that a comb filter that attenuates all frequency components for a frame that does not include speech components is generated. .

  In FIG. 7, the FFT unit 103 performs FFT on the audio signal output from the windowing unit 102, and outputs the audio spectrum signal to the frequency division unit 104 and the section determination unit 701.

  The section determination unit 701 determines whether the speech spectrum output from the FFT unit 103 includes speech and outputs the determination result to the comb filter reset unit 702.

  When it is determined that the speech spectrum is only a noise component that does not include a speech component based on the determination result output from the section determination unit 701, the comb filter reset unit 702 causes the comb filter generation unit 107 to combine all frequency components. Outputs instructions to turn off the filter.

  The comb filter generation unit 107 generates a comb filter that emphasizes pitch harmonics based on the presence or absence of a voice component in each frequency component, and outputs the comb filter to the attenuation coefficient calculation unit 108. Further, the comb filter generation unit 107 generates and attenuates a comb filter that is turned off for all frequency components when it is determined that the audio spectrum includes only a noise component that does not include the audio component in accordance with an instruction from the comb filter reset unit 702. It outputs to the coefficient calculation part 108.

  As described above, according to the audio processing device according to the fifth embodiment of the present invention, the frame that does not include the audio component is attenuated with the entire frequency component, and the noise is cut in the entire band in the signal interval that does not include the audio. Thus, it is possible to prevent the occurrence of noise due to the voice suppression process, and thus voice enhancement with less voice distortion can be performed.

  Note that Embodiment 5 can be combined with Embodiment 2 or Embodiment 3.

  That is, the effects of the second embodiment can be obtained by adding the noise section determination unit 401 and the noise base tracking unit 402 to the speech processing device of FIG. 7, and the musical noise suppressing unit 501 and the speech processing device of FIG. The effect of the third embodiment can be obtained by adding the comb filter correction unit 502.

  Further, the fifth embodiment can be combined with the fourth embodiment. That is, the effect of the fourth embodiment can be obtained by adding the average value calculation unit 601 to the speech processing apparatus of FIG.

  In this case, the frequency division unit 104 divides the voice spectrum output from the FFT unit 103 into frequency components indicating the voice spectrum divided in predetermined frequency units, and the voice spectrum is classified into voice non-voice for each frequency component. Unit 106, multiplier 109, and average value calculator 601.

  When the difference between the average value of the speech spectrum signal output from the average value calculation unit 601 and the noise base value output from the noise base estimation unit 105 is equal to or greater than a predetermined threshold, the speech non-speech identification unit 106 If the difference is smaller than a predetermined threshold value, it is determined that the sound is a silent part including only noise, and the determination result is determined as a noise base estimation unit 105 and a comb filter generation unit 107. Output to.

(Embodiment 6)
FIG. 8 is a block diagram of an example of the configuration of the speech processing apparatus according to the sixth embodiment. However, the same components as those in FIG. 1 are denoted by the same reference numerals as those in FIG.

  The voice processing apparatus of FIG. 8 includes a voice pitch period estimation unit 801 and a voice pitch restoration unit 802, and compensates for pitch harmonic information that is lost as a result of being judged as noise in a frequency region where it is difficult to judge voice and noise. Different from FIG.

  In FIG. 8, the frequency dividing unit 104 divides the audio spectrum output from the FFT unit 103 into frequency components indicating the audio spectrum divided in predetermined frequency units, and noise-based estimation of the audio spectrum for each frequency component. Unit 105, speech non-speech identification unit 106, multiplication unit 109, speech pitch period estimation unit 801, and speech pitch restoration unit 802.

  The comb filter generation unit 107 generates a comb filter that emphasizes pitch harmonics based on the presence or absence of a voice component in each frequency component, and outputs the generated comb filter to the voice pitch period estimation unit 801 and the voice pitch restoration unit 802.

  The voice pitch period estimation unit 801 estimates the pitch period from the comb filter output from the comb filter generation unit 107 and the voice spectrum output from the frequency division unit 104, and outputs the estimation result to the voice pitch restoration unit 802.

For example, one frequency component is turned off without being continuously turned on in the generated comb filter. Next, a pitch period estimation comb filter is generated by extracting two frequency components with high power from the comb filter, and the pitch period is obtained from the following autocorrelation function equation (7).

  Here, PITCH (k) represents the state of the pitch period estimation comb filter, k1 represents the upper limit of the frequency, τ represents the pitch period, and τ is a value from 0 to τ1, which is the maximum pitch period. Take.

  Τ having the maximum value of γ (τ) in Expression (7) is obtained as the pitch period. Actually, since the shape of the frequency pitch tends to be unclear in the high frequency region, an intermediate frequency value is used for k1. For example, set k1 = 2 kHz. In addition, by setting the values that PITCH (k) can take to be 0 and 1, the calculation of Expression (7) can be easily performed.

  The voice pitch restoration unit 802 corrects the comb filter based on the estimation result output from the voice pitch period estimation unit 801 and outputs the result to the attenuation coefficient calculation unit 108. Specifically, based on the estimated pitch period information, the pitch is compensated for every fixed frequency component, or the width of the comb-shaped band that is a continuation of the frequency components that are turned on in the comb filter for each pitch period To repair the pitch harmonic structure.

  The attenuation coefficient calculation unit 108 multiplies the comb filter output from the audio pitch restoration unit 802 by an attenuation coefficient based on the frequency characteristics, sets the attenuation coefficient of the input signal for each frequency component, and sets each frequency component. The attenuation coefficient is output to the multiplier 109.

  FIG. 9 shows an example of comb filter restoration in the speech processing apparatus according to the present embodiment. In FIG. 9, the vertical axis represents the attenuation, and the horizontal axis represents the frequency component. Specifically, the horizontal axis has 256 frequency components and represents a region from 0 kHz to 4 kHz.

  C1 indicates a generated comb filter, C2 indicates a comb filter obtained by correcting the pitch of the comb filter C1, and C3 indicates a comb filter whose pitch width is corrected by the comb filter C2.

  In the comb filter C1, pitch information is lost in frequency components from 100 to 140. The voice pitch repair unit 802 supplements the pitch information in the frequency components from 100 to 140 of the comb filter C1 based on the pitch period information estimated by the voice pitch period estimation unit 801. Thereby, the comb filter C2 is obtained.

  Next, the audio pitch restoration unit 802 corrects the pitch harmonic width of the comb filter C2 based on the audio spectrum output from the frequency division unit 104. Thereby, the comb filter C3 is obtained.

  As described above, according to the speech processing device according to Embodiment 6 of the present invention, the speech that is close to the original speech is estimated by estimating the pitch period information and compensating for the lost pitch harmonic information that is determined as noise. In this state, it is possible to perform speech enhancement with little speech distortion.

  Note that Embodiment 6 can be combined with Embodiment 2 or Embodiment 5.

  That is, the effects of the second embodiment can be obtained by adding the noise section determination unit 401 and the noise base tracking unit 402 to the voice processing apparatus of FIG. 8, and the section determination unit 701 and the voice processing apparatus of FIG. By adding the comb filter reset unit 702, the effect of the fifth embodiment can be obtained.

  Further, the sixth embodiment can be combined with the third embodiment. That is, the effects of the third embodiment can be obtained by adding the musical noise suppression unit 501 and the comb filter correction unit 502 to the speech processing apparatus of FIG.

  In this case, the musical noise suppression unit 501 is ON among the frequency components of the comb filter output from the comb filter generation unit 107, that is, the number of states in which the signal is output without being attenuated is equal to or less than a certain threshold value. Then, it is determined that sudden noise is included in the frame, and the determination result is output to the voice pitch period estimation unit 801.

  The comb filter correction unit 502 performs correction to prevent the generation of musical noise in the comb filter based on the comb filter generation result output from the comb filter generation unit 107 in which sudden noise is included in the frame from the voice pitch restoration unit 802. The comb filter is output to the attenuation coefficient calculation unit 108.

  Further, the sixth embodiment can be combined with the fourth embodiment. That is, the effect of the fourth embodiment can be obtained by adding the average value calculation unit 601 to the speech processing apparatus of FIG.

  In this case, the frequency division unit 104 divides the voice spectrum output from the FFT unit 103 into frequency components indicating the voice spectrum divided in predetermined frequency units, and the voice spectrum is classified into voice non-voice for each frequency component. Unit 106, multiplier 109, and average value calculator 601.

  When the difference between the average value of the speech spectrum signal output from the average value calculation unit 601 and the noise base value output from the noise base estimation unit 105 is equal to or greater than a predetermined threshold, the speech non-speech identification unit 106 If the difference is smaller than a predetermined threshold value, it is determined that the sound is a silent part including only noise, and the determination result is determined as a noise base estimation unit 105 and a comb filter generation unit 107. Output to.

(Embodiment 7)
FIG. 10 is a block diagram of an example of the configuration of the speech processing apparatus according to the seventh embodiment. However, the same components as those in FIGS. 1 and 4 are denoted by the same reference numerals as those in FIGS. 1 and 4, and detailed description thereof is omitted. The speech processing apparatus in FIG. 10 includes an automatic threshold adjustment unit 1001 and is different from FIG. 1 or FIG. 4 in that the speech identification threshold is adjusted according to the type of noise.

  In FIG. 10, the comb filter generation unit 107 generates a comb filter that emphasizes pitch harmonics based on the presence or absence of a voice component in each frequency component, and outputs the comb filter to the threshold automatic adjustment unit 1001.

  The noise section determination unit 401 calculates the signal power and moving average value for each frame from the speech spectrum output from the FFT unit 103, and determines whether or not the frame includes speech from the rate of change of the input signal power. The discrimination result is output to the threshold automatic adjustment unit 1001.

  The threshold automatic adjustment unit 1001 determines that the speech non-speech discrimination unit 106 is based on the comb filter output from the comb filter generation unit 107 when the speech signal is not included in the frame from the discrimination result output from the noise section discrimination unit 401. Change the threshold value.

Specifically, the sum total of the number of frequency components in the on state of the comb filter generated using the following equation (8) is calculated.

  When this sum exceeds a predetermined upper limit value, an instruction to increase the threshold value of the voice non-speech identification unit 106 is given. When this sum becomes smaller than a predetermined lower limit value, the threshold value of the voice non-speech discrimination unit 106 is decreased. To the voice non-voice identification unit 106.

  Here, n1 is a number that identifies a frame that has been processed in the past, and n2 is a number that identifies a frame that has been processed.

  For example, when noise with small amplitude variation is included in a frame, the threshold value for voice non-voice identification is set low, and when noise with large amplitude variation is included in a frame, the threshold value for voice non-voice identification is set high.

  Thus, according to the speech processing device according to the embodiment of the present invention, speech non-speech in the speech spectrum is based on the number of frequency components that are erroneously determined to contain speech in a frame that does not include speech. By changing the threshold value used for speech identification, speech corresponding to the type of noise can be determined, and speech enhancement with less speech distortion can be performed.

  Note that Embodiment 7 can be combined with Embodiment 2 or Embodiment 3.

  In other words, the effects of the second embodiment can be obtained by adding the noise section determination unit 401 and the noise base tracking unit 402 to the speech processing device of FIG. 10, and the musical noise suppression unit 501 and the noise processing unit of FIG. The effect of the third embodiment can be obtained by adding the comb filter correction unit 502.

  The seventh embodiment can be combined with the fourth embodiment. That is, the effect of the fourth embodiment can be obtained by adding the average value calculation unit 601 to the speech processing apparatus of FIG.

  In this case, the frequency division unit 104 divides the voice spectrum output from the FFT unit 103 into frequency components indicating the voice spectrum divided in predetermined frequency units, and the voice spectrum is classified into voice non-voice for each frequency component. Unit 106, multiplier 109, and average value calculator 601.

  When the difference between the average value of the speech spectrum signal output from the average value calculation unit 601 and the noise base value output from the noise base estimation unit 105 is equal to or greater than a predetermined threshold, the speech non-speech identification unit 106 If the difference is smaller than a predetermined threshold value, it is determined that the sound is a silent part including only noise, and the determination result is determined as a noise base estimation unit 105 and a comb filter generation unit 107. Output to.

  Further, Embodiment 7 can be combined with Embodiment 5 or Embodiment 6. That is, the effect of the fifth embodiment can be obtained by adding the section determination unit 701 and the comb filter reset unit 702 to the speech processing device of FIG. 10, and the speech pitch period estimation unit 801 is added to the speech processing device of FIG. In addition, the effect of the sixth embodiment can be obtained by adding the audio pitch restoration unit 802.

(Embodiment 8)
FIG. 11 is a block diagram of an example of the configuration of the speech processing apparatus according to the eighth embodiment. However, the same components as those in FIG. 1 are denoted by the same reference numerals as those in FIG.

  The speech processing apparatus in FIG. 11 includes a noise base estimation unit 1101, a first speech non-speech identification unit 1102, a second speech non-speech discrimination unit 1103, a speech pitch estimation unit 1104, and a first comb filter generation unit 1105. A second comb filter generation unit 1106, a voice pitch restoration unit 1107, a comb filter correction unit 1108, and a voice separation coefficient calculation unit 1109, and a noise base used for creating a comb filter and a pitch harmonic structure restoration. 1 is different from the speech processing apparatus of FIG. 1 in that the noise base used for is generated under different conditions.

  In FIG. 11, the frequency division unit 104 divides the speech spectrum output from the FFT unit 103 into frequency components, and converts the speech spectrum for each frequency component into a noise base estimation unit 1101, a first speech non-speech identification unit 1102, The data is output to the second voice / non-voice identification unit 1103 and the voice pitch estimation unit 1104.

  When the determination result that the speech component is included in the frame is output from the first speech non-speech identification unit 1102, the noise base estimation unit 1101 outputs the noise base estimated in the past to the first speech non-speech discrimination unit 1102. To do. Further, when the determination result that the speech component is included in the frame is output from the second speech non-speech identification unit 1103, the noise base estimation unit 1101 determines the noise base estimated in the past as the second speech non-speech discrimination unit 1103. Output to.

  In addition, the noise base estimation unit 1101 outputs from the frequency division unit 104 when a determination result that a frame does not include a speech component is output from the first speech non-speech discrimination unit 1102 or the second speech non-speech discrimination unit 1103. Calculate a moving average value that represents the average amount of short-time power spectrum and spectrum change for each frequency component of the acquired voice spectrum, take the moving average value calculated in the past and the weighted average value of the power spectrum, and create a new moving average Calculate the value.

Specifically, the noise base estimation unit 1101 estimates the noise base in each frequency component using the equation (9) or the equation (10), and the first speech non-speech discrimination unit 1102 or the second speech non-speech discrimination unit. 1103.

Here, n is a number for specifying a frame to be processed, k is a number for specifying a frequency component, and τ is a delay time. S ² _f (n, k) represents the power spectrum of the input audio signal, P _base (n, k) represents a noise-based moving average value, and α (k) represents a moving average coefficient.

  When the power spectrum of the input speech signal is equal to or less than the result of multiplying the threshold value for determining speech and noise by the power spectrum of the previously input speech signal, the noise base estimation unit 1101 obtains from Equation (9). Output the noise base. Further, when the power spectrum of the input audio signal is larger than the multiplication result of the threshold value for discriminating audio and noise and the power spectrum of the previously input audio signal, the noise base estimation unit 1101 obtains the equation (10). The obtained noise base is output.

  When the difference between the speech spectrum signal output from the frequency division unit 104 and the noise base value output from the noise base estimation unit 1101 is equal to or greater than a predetermined first threshold, the first speech non-speech identification unit 1102 In other cases, it is determined that the sound part is a silent part including only noise.

  In the first speech non-speech identification unit 1102, a second speech non-speech discrimination unit 1103, which will be described later, is used by the first comb filter generation unit 1105 to generate a filter that extracts as much pitch harmonic information as possible. Set to a value lower than the second threshold value to be used. Then, the first voice / non-voice identification unit 1102 outputs the determination result to the first comb filter generation unit 1105.

  When the difference between the speech spectrum signal output from the frequency division unit 104 and the noise base value output from the noise base estimation unit 1101 is equal to or greater than a predetermined second threshold, the second speech non-speech identification unit 1103 In other cases, it is determined that the sound part is a silent part including only noise. Then, the second voice / non-voice identification unit 1103 outputs the determination result to the second comb filter generation unit 1106.

  First comb filter generation section 1105 generates a first comb filter that emphasizes pitch harmonics based on the presence or absence of audio components in each frequency component, and outputs the first comb filter to comb filter correction section 1108.

Specifically, in the first speech non-speech identification unit 1102, the power spectrum of the input speech signal is equal to or greater than the result of multiplying the first threshold value for distinguishing speech and noise by the power spectrum of the input speech signal. In some cases, that is, when Expression (11) is satisfied, the first comb filter generation unit 1105 sets the filter value of the frequency component to “1”.

Further, when the power spectrum of the input voice signal is smaller than the multiplication result of the first threshold value for determining voice and noise and the power spectrum of the input voice signal in the first voice / non-voice identification unit 1102, that is, If the expression (12) is satisfied, the first comb filter generation unit 1105 sets the comb filter value of the frequency component to “0”.

Here, k is a number that specifies the frequency component, and satisfies the value of the following equation (13). HB indicates the number of data points when fast Fourier transform is performed on an audio signal.

  Second comb filter generation section 1106 generates a second comb filter that emphasizes pitch harmonics based on the presence or absence of a voice component in each frequency component, and outputs the second comb filter to voice pitch restoration section 1107.

Specifically, in the second speech non-speech discrimination unit 1103, the power spectrum of the input speech signal is equal to or greater than the result of multiplying the second threshold value for distinguishing speech and noise by the power spectrum of the input speech signal. In some cases, that is, when Expression (11) is satisfied, the second comb filter generation unit 1106 sets the filter value of the frequency component to “1”.

Further, in the second speech non-speech identifying unit 1103, when the power spectrum of the input speech signal is smaller than the multiplication result of the second threshold value for distinguishing speech and noise and the power spectrum of the input speech signal, When the expression (12) is satisfied, the second comb filter generation unit 1106 sets the filter value of the frequency component to “0”.

  The voice pitch estimation unit 1104 estimates the pitch period from the voice spectrum output from the frequency division unit 104 and outputs the estimation result to the voice pitch restoration unit 1107.

ここで、ＣＯＭＢ_low(k)は、第一コムフィルタ生成部１１０５において生成された第一コムフィルタを示す。ｋ１は、周波数の上限値を示す。また、τは、ピッチの周期を示し、「０」からピッチの最大周期までの値をとる。 For example, the voice pitch estimation unit 1104 obtains the pitch period using the following autocorrelation function formula (14) for the voice spectrum power in the passband of the generated comb filter.

Here, COMB_low (k) indicates the first comb filter generated by the first comb filter generation unit 1105. k1 represents an upper limit value of the frequency. Also, τ indicates the pitch period, and takes a value from “0” to the maximum pitch period.

  Then, the voice pitch estimation unit 1104 obtains τ for which γ (τ) has the maximum value as the voice pitch period. In actual processing, the shape of pitch harmonics is often unclear in the high frequency region. Therefore, the intermediate frequency value is used for k1, and the pitch period of the low frequency side half of the frequency region of the audio signal is used. Estimate. For example, the voice pitch estimation unit 1104 estimates the voice pitch period by setting k1 = 2 kHz.

  The voice pitch restoration unit 1107 corrects the second comb filter based on the estimation result output from the voice pitch estimation unit 1104 and outputs the result to the comb filter correction unit 1108.

  Hereinafter, a specific operation of the audio pitch restoration unit 1107 will be described with reference to the drawings. 12, FIG. 13, FIG. 14 and FIG. 15 are diagrams showing examples of comb filters.

  The voice pitch restoration unit 1107 extracts the peak of the pass region of the second comb filter and generates a pitch reference comb filter. The comb filter in FIG. 12 is an example of the second comb filter generated by the second comb filter generation unit 1106. Moreover, the comb filter of FIG. 13 is an example of a pitch reference comb filter. In the comb filter of FIG. 13, only peak information is extracted from the comb filter of FIG.

  Then, the voice pitch restoration unit 1107 calculates the peak-to-peak interval of the pitch reference comb filter, and if the peak-to-peak interval exceeds a predetermined threshold, for example, a value that is 15 times the pitch period, the voice pitch The pitch insertion comb filter is generated by inserting the missing pitch from the pitch estimation result of the estimation unit 1104. The comb filter in FIG. 14 is an example of a pitch insertion comb filter. In the comb filter of FIG. 14, peaks are inserted in the vicinity of 50 kHz to 100 kHz and from 200 kHz to 250 kHz.

  Then, the audio pitch repair unit 1107 generates a pitch repair comb filter by expanding the width of the peak of the pass region of the pitch insertion comb filter according to the pitch value, and outputs the generated pitch repair comb filter to the comb filter correction unit 1108. The comb filter of FIG. 15 is an example of a pitch repair comb filter. In the comb filter of FIG. 15, information on the width of the pass region is added to the pitch insertion comb filter of FIG.

  The comb filter correction unit 1108 corrects the first comb filter generated in the first comb filter generation unit 1105 using the pitch repair comb filter generated in the voice pitch repair unit 1107, and converts the corrected comb filter into a voice separation coefficient. The result is output to the calculation unit 1109.

  Specifically, the comb filter correction unit 1108 compares the pass regions of the pitch restoration comb filter and the first comb filter, and sets a portion that is a pass region in both comb filters as a pass region. A comb filter is generated as a blocking region for attenuating the signal.

  An example of comb filter correction will be shown below. 16, FIG. 17, and FIG. 18 are diagrams illustrating examples of comb filters. The comb filter in FIG. 16 is a first comb filter generated by the first comb filter generation unit 1105. Further, the comb filter of FIG. 17 is a pitch repair comb filter generated in the voice pitch repair unit 1107. FIG. 18 is an example of a comb filter modified by the comb filter modification unit 1108.

  The speech separation coefficient calculation unit 1109 multiplies the comb filter modified by the comb filter modification unit 1108 by a separation coefficient based on the frequency characteristics, calculates a separation coefficient of the input signal for each frequency component, and outputs it to the multiplication unit 109. To do.

For example, if the value of the comb filter COMB_res (k) corrected by the comb filter correction unit 1108 is 1, that is, the pass region, in the number k that identifies a certain frequency component, the speech separation coefficient calculation unit 1109 has a dispersion coefficient sps. Let (k) be 1. Also, when the value of the comb filter COMB_res (k) is 0, that is, the blocking region, the speech separation coefficient calculation unit 1109 calculates the dispersion coefficient sps (k) from the following equation (15).

  Here, gc is a constant, k is a number for specifying a frequency component, and HB is an FFT transform length, that is, the number of data to be subjected to fast Fourier transform.

  The multiplication unit 109 multiplies the speech spectrum output from the frequency division unit 104 by the attenuation coefficient output from the speech separation coefficient calculation unit 1109 in units of frequency components. Then, the spectrum obtained as a result of multiplication is output to frequency synthesizer 110.

  Thus, according to the speech processing apparatus of the present embodiment, a large amount of speech information is extracted by generating a noise base used for creating a comb filter and a noise base used for pitch harmonic structure restoration under different conditions. In addition, it is possible to generate a comb filter that is not easily affected by noise information and to accurately restore the pitch harmonic structure.

  Specifically, according to the speech processing apparatus of the present embodiment, the pitch estimated to be missing is inserted by reflecting the estimation result of the pitch period based on the second comb filter with strict conditions for determining speech. By repairing the pitch harmonic structure of the comb filter, it is possible to reduce voice distortion due to the lack of pitch harmonics.

  Moreover, according to the speech processing apparatus of the present embodiment, the pitch harmonic structure can be accurately restored by adjusting the pitch width of the comb filter from the estimation result of the pitch period. The overlapping part of the comb filter pass region created by strict judgment of speech as the comb filter's pitch harmonic structure and the comb filter pass region created loosely as speech is defined as the pass region. By creating a comb filter having a blocking region other than the above, it is possible to reduce the influence of an error in estimating the pitch period, and it is possible to accurately restore the pitch harmonic structure.

  Note that the speech processing apparatus according to the present embodiment calculates the speech separation coefficient in the blocking region of the comb filter by multiplying the speech spectrum by the separation factor, and calculates the speech separation coefficient in the passing region of the comb filter from the speech spectrum. It can also be calculated by subtracting the base.

For example, when the value of the comb filter COMB_res (k) is 0, that is, the blocking region, the speech separation coefficient calculation unit 1109 calculates the dispersion coefficient sps (k) from the following equation (16).

Here, P _max (n) represents the maximum value of P _base (n, k) in the frequency component k within a predetermined range. In Expression (16), the noise base estimation value is normalized for each frame, and the reciprocal is used as the separation coefficient.

Then, when the value of the comb filter COMB_res (k) is 1, that is, the pass region, the dispersion coefficient seps (k) is calculated from the following equation (17).

Here, γ is a coefficient indicating the amount by which the noise base is subtracted, and P _max (n) indicates the maximum value of P _base (n, k) in the frequency component k within a predetermined range.

  As described above, the speech processing apparatus according to the present embodiment multiplies the inhibition region of the pitch-combined comb filter by the separation coefficient calculated from the noise-based information, so that optimum separation is achieved even for different noise characteristics. Coefficients can be calculated, and speech enhancement corresponding to noise characteristics can be performed. Also, the speech processing apparatus according to the present embodiment performs speech enhancement with less speech distortion by multiplying the passing region of the comb filter that has undergone pitch correction by the separation coefficient calculated by subtracting the noise base from the speech spectrum. be able to.

  Further, this embodiment can be combined with Embodiment 2. In other words, the effects of the second embodiment can be obtained by adding the noise section determination unit 401 and the noise base tracking unit 402 to the speech processing apparatus of FIG.

(Embodiment 9)
FIG. 19 is a block diagram of an example of the configuration of the speech processing apparatus according to the ninth embodiment. However, the same components as those in FIGS. 1 and 11 are denoted by the same reference numerals as those in FIGS. 1 and 11, and detailed description thereof is omitted.

  19 includes an SNR calculator 1901 and an audio noise frame detector 1902, calculates an SNR (Signal Noise Ratio) of the audio signal, and calculates an audio frame or noise from the audio signal in units of frames from the SNR. It differs from FIG. 1 or FIG. 11 in that the frame period is detected and the pitch period is estimated only for the audio frame.

  In FIG. 19, the frequency dividing unit 104 divides the audio spectrum output from the FFT unit 103 into frequency components, and converts the audio spectrum for each frequency component into the noise base estimation unit 105 and the first audio non-speech identification unit 1102. And the second voice / non-voice identification unit 1103, the multiplication unit 109, and the SNR calculation unit 1901.

  First comb filter generation section 1105 generates a first comb filter that emphasizes pitch harmonics based on the presence or absence of audio components in each frequency component, and outputs the first comb filter to comb filter correction section 1108 and SNR calculation section 1901.

The SNR calculation unit 1901 calculates the SNR of the audio signal from the audio spectrum output from the frequency division unit 104 and the first comb filter output from the first comb filter generation unit 1105, and outputs the SNR to the audio noise frame detection unit 1902. . For example, the SNR calculator 1901 calculates the SNR using the following equation (18).

  Here, COMB_low (k) indicates the first comb filter. K represents a frequency component, and takes a value of 0 or more and smaller than half of the number of data points when the fast Fourier transform is performed on the audio signal.

  The voice noise frame detection unit 1902 determines whether the input signal is a voice signal or a noise signal in units of frames from the SNR output from the SNR calculation unit 1901 and outputs the determination result to the voice pitch estimation unit 1903. Specifically, when the SNR is greater than a predetermined threshold, the audio noise frame detection unit 1902 determines that the input signal is an audio signal (audio frame), and the number of frames having an SNR equal to or less than the predetermined threshold is equal to or greater than a predetermined number. When it occurs continuously, the input signal is determined as a noise signal (noise frame).

  FIG. 20 shows an example in which the speech / noise determination operation of the speech noise frame detection unit 1902 is expressed by a program. FIG. 20 is a diagram illustrating an example of a speech noise determination program of the speech processing device according to the present embodiment. In the program of FIG. 20, when 10 or more frames having an SNR equal to or less than a predetermined threshold occur continuously, the input signal is determined as a noise signal (noise frame).

  When the speech noise frame detection unit 1902 determines that the speech noise frame detection unit 1902 is a speech frame, the speech pitch estimation unit 1903 estimates the pitch period from the speech spectrum output from the frequency division unit 104 and outputs the estimation result to the speech pitch restoration unit 1107. The operation of pitch period estimation is the same as that of speech pitch estimation section 1104 of the eighth embodiment.

  The voice pitch restoration unit 1107 corrects the second comb filter based on the estimation result output from the voice pitch estimation unit 1903 and outputs the result to the comb filter correction unit 1108.

  Thus, according to the speech processing apparatus of the present embodiment, the ratio of the sum of the power of the speech spectrum corresponding to the pass region of the comb filter and the sum of the power of the speech spectrum corresponding to the blocking region of the comb filter is calculated. By obtaining the SNR and estimating the pitch period using only frames whose SNR is equal to or greater than a predetermined threshold, it is possible to reduce pitch period estimation errors due to noise, and to perform speech enhancement with less audio distortion Can do.

  In addition, although the speech processing apparatus of this Embodiment calculates SNR from the 1st comb filter, you may calculate SNR using a 2nd comb filter. In this case, the second comb filter generation unit 1106 outputs the created second comb filter to the SNR calculation unit 1901. Then, the SNR calculator 1901 calculates the SNR of the audio signal from the audio spectrum output from the frequency divider 104 and the second comb filter, and outputs the SNR to the audio noise frame detector 1902.

(Embodiment 10)
FIG. 21 is a block diagram of an example of the configuration of the speech processing apparatus according to the tenth embodiment. However, the same components as those in FIGS. 1 and 11 are denoted by the same reference numerals as those in FIGS. 1 and 11, and detailed description thereof is omitted. 21 includes a first comb filter generation unit 2101, a first musical noise suppression unit 2102, a second comb filter generation unit 2103, and a second musical noise suppression unit 2104. It differs from FIG. 1 or FIG. 11 in that the generation of musical noise is determined from the generation results of the filter and the second comb filter.

  In FIG. 21, the first speech non-speech identification unit 1102 has a difference between the speech spectrum signal output from the frequency division unit 104 and the noise base value output from the noise base estimation unit 1101 equal to or greater than a predetermined first threshold. In the case, it is determined that the sound part includes a sound component. In other cases, it is determined that the sound part includes only a noise that does not include a sound component.

  In the first speech non-speech identification unit 1102, the second speech non-speech discrimination unit 1103, which will be described later, is used by the first comb filter generation unit 2101 to generate a filter that extracts as much pitch harmonic information as possible. Set to a value lower than the second threshold value to be used. Then, the first voice / non-voice identification unit 1102 outputs the determination result to the first comb filter generation unit 2101.

  When the difference between the speech spectrum signal output from the frequency division unit 104 and the noise base value output from the noise base estimation unit 1101 is equal to or greater than a predetermined second threshold, the second speech non-speech identification unit 1103 In other cases, it is determined that the sound part is a silent part including only noise. Then, the second voice / non-voice identification unit 1103 outputs the determination result to the second comb filter generation unit 2103.

  First comb filter generation section 2101 generates a first comb filter that emphasizes pitch harmonics based on the presence or absence of audio components in each frequency component, and outputs the first comb filter to first musical noise suppression section 2102. The specific operation of the first comb filter generation is the same as that of the first comb filter generation unit 1105 of the eighth embodiment. Then, the first comb filter generation unit 2101 outputs the first comb filter corrected by the first musical noise suppression unit 2102 to the comb filter correction unit 1108.

The first musical noise suppression unit 2102 turns on sudden noise in a frame when the frequency components of the first comb filter are on, that is, when the number of states to be output without attenuation is equal to or less than a certain threshold value. Is determined to be included. For example, if the number of frequency components turned on by the comb filter is calculated using the following equation (5), and COMB_SUM (n) is smaller than a certain threshold value (for example, 10), musical noise is generated. to decide.

  Then, the first musical noise suppression unit 2102 sets the state of all frequency components of the comb filter to OFF, that is, sets the signal to be attenuated and outputs it, and outputs the comb filter to the first comb filter generation unit 2101.

  Second comb filter generation section 2103 generates a second comb filter that emphasizes pitch harmonics based on the presence or absence of audio components in each frequency component, and outputs the second comb filter to second musical noise suppression section 2104. The specific operation of generating the second comb filter is the same as that of the second comb filter generation unit 1106 of the eighth embodiment. Then, the second comb filter generation unit 2103 outputs the second comb filter corrected by the second musical noise suppression unit 2104 to the audio pitch restoration unit 1107.

  The second musical noise suppression unit 2104 turns on sudden noise in a frame when the frequency components of the first comb filter are on, that is, when the number of states in which signals are output without attenuation is equal to or less than a certain threshold. Is determined to be included.

For example, if the number of frequency components turned on by the comb filter is calculated using the following equation (5), and COMB_SUM (n) is smaller than a certain threshold value (for example, 10), musical noise is generated. to decide.

  Then, the second musical noise suppression unit 2104 sets all the frequency components of the comb filter to the off state, that is, sets the signal to be attenuated and outputs it, and outputs the comb filter to the second comb filter generation unit 2103.

  The voice pitch restoration unit 1107 corrects the second comb filter output from the second comb filter generation unit 2103 based on the estimation result output from the voice pitch estimation unit 1104, and outputs the result to the comb filter correction unit 1108.

  The comb filter correction unit 1108 corrects the first comb filter generated in the first comb filter generation unit 2101 using the pitch correction comb filter generated in the voice pitch correction unit 1107, and the corrected comb filter is converted into a voice separation coefficient. The result is output to the calculation unit 1109.

  As described above, according to the audio processing device of the present embodiment, it is possible to prevent the noise from being erroneously determined as an audio signal by determining the occurrence of musical noise from the generation results of the first comb filter and the second comb filter. Therefore, it is possible to perform speech enhancement with little speech distortion.

(Embodiment 11)
FIG. 22 is a block diagram of an example of the configuration of the speech processing apparatus according to the eleventh embodiment. However, the same components as those in FIGS. 1 and 11 are denoted by the same reference numerals as those in FIGS. 1 and 11, and detailed description thereof is omitted. The speech processing apparatus of FIG. 22 includes an average value calculation unit 2201 and is different from FIG. 1 or FIG. 11 in that the average value of speech spectrum power is obtained in units of frequency components.

  In FIG. 22, the frequency dividing unit 104 divides the speech spectrum output from the FFT unit 103 into frequency components, and converts the speech spectrum for each frequency component into a noise-based estimation unit 1101 and a first speech non-speech identifying unit 1102. To the multiplication unit 109 and the average value calculation unit 2201.

  The average value calculation unit 2201 takes the average value with the frequency components in the vicinity and the average value with the previously processed frame for the power of the audio spectrum output from the frequency division unit 104, and uses the obtained average value as the second value. Output to the voice / non-voice identification unit 1103.

Specifically, the average value of the speech spectrum is calculated using the following equation (19).

  Here, k1 and k2 indicate frequency components, and k1 <k <k2. n1 indicates a number indicating a frame that has been processed in the past, and n indicates a number indicating a frame that has been processed.

  The second speech non-speech identification unit 1103 has a difference between the average value of the speech spectrum signal output from the average value calculation unit 2201 and the noise base value output from the noise base estimation unit 1101 equal to or greater than a predetermined second threshold. In the case, it is determined that the sound part includes a sound component. In other cases, it is determined that the sound part includes only a noise that does not include a sound component. Then, the second voice / non-voice identification unit 1103 outputs the determination result to the second comb filter generation unit 1106.

  As described above, according to the speech processing apparatus according to Embodiment 11 of the present invention, the average power value of the speech spectrum in each frequency component or the average power value of the frame that has been processed in the past and the frame that has been processed is obtained. Thus, the influence of the sudden noise component is reduced, and the second comb filter that extracts only the voice information can be generated more accurately.

(Embodiment 12)
FIG. 23 is a block diagram of an example of the configuration of the speech processing apparatus according to the twelfth embodiment. However, the same components as those in FIGS. 1, 11, and 19 are denoted by the same reference numerals as those in FIGS. 1, 11, and 19, and detailed description thereof is omitted. The speech processing apparatus of FIG. 23 includes a comb filter reset unit 2301 and generates a comb filter that attenuates all frequency components for a frame that does not include speech components, as shown in FIG. 1, FIG. 11, or FIG. Different.

  In FIG. 23, the speech noise frame detection unit 1902 determines whether the input signal is a speech signal or a noise signal in units of frames from the SNR output from the SNR calculation unit 1901 and outputs the determination result to the speech pitch estimation unit 1104.

  Specifically, when the SNR is greater than a predetermined threshold, the audio noise frame detection unit 1902 determines that the input signal is an audio signal (audio frame), and the number of frames having an SNR equal to or less than the predetermined threshold is equal to or greater than a predetermined number. When it occurs continuously, the input signal is determined as a noise signal (noise frame). Then, the speech noise frame detection unit 1902 outputs the determination result to the speech pitch estimation unit 1104 and the comb filter reset unit 2301.

  When the comb filter reset unit 2301 determines based on the determination result output from the voice noise frame detection unit 1902 that the voice spectrum includes only a noise component that does not include a voice component, the comb filter correction unit 1108 sends all frequency components to the comb filter correction unit 1108. Outputs an instruction to turn off the comb filter.

  The comb filter correction unit 1108 corrects the first comb filter generated in the first comb filter generation unit 1105 using the pitch repair comb filter generated in the voice pitch repair unit 1107, and converts the corrected comb filter into a voice separation coefficient. The result is output to the calculation unit 1109.

  Further, the comb filter correction unit 1108 generates a first comb filter that is turned off for all frequency components when it is determined that the voice spectrum is only a noise component that does not include a voice component in accordance with an instruction from the comb filter reset unit 2301. To the speech separation coefficient calculation unit 1109.

  As described above, according to the speech processing apparatus of the present embodiment, speech suppression is performed by performing attenuation on all frequency components in a frame that does not include speech components, and cutting noise on all bands in signal sections that do not include speech. Since it is possible to prevent the occurrence of noise due to processing, it is possible to perform voice enhancement with less voice distortion.

(Embodiment 13)
FIG. 24 is a block diagram of an example of the configuration of the speech processing apparatus according to the thirteenth embodiment. However, the same components as those in FIG. 1 are denoted by the same reference numerals as those in FIG.

  The speech processing apparatus of FIG. 24 includes a noise separation comb filter generation unit 2401, a noise separation coefficient calculation unit 2402, a multiplication unit 2403, and a noise frequency synthesis unit 2404, and the speech non-speech of the spectrum signal in units of frequency components. The frequency characteristics are attenuated in units of frequency components, the frequency characteristics are attenuated based on the discrimination results, a comb filter that extracts only noise components by obtaining accurate pitch information is created, and noise characteristics are extracted. Different from the speech processing apparatus.

  The speech non-speech identification unit 106, when the difference between the speech spectrum signal output from the frequency division unit 104 and the noise base value output from the noise base estimation unit 105 is greater than or equal to a predetermined threshold value, It is determined to be a part, and otherwise, it is determined to be a silent part of only noise that does not include a voice component. Then, the speech non-speech identification unit 106 outputs the determination result to the noise base estimation unit 105 and the noise separation comb filter generation unit 2401.

  The noise separation comb filter generation unit 2401 generates a comb filter that emphasizes pitch harmonics based on the presence or absence of a voice component in each frequency component, and outputs this comb filter to the noise separation coefficient calculation unit 2402.

Specifically, in the speech non-speech identification unit 106, when the power spectrum of the input speech signal is equal to or greater than the multiplication result of the first threshold value for determining speech and noise and the power spectrum of the input speech signal. That is, when Expression (20) is satisfied, the noise separation comb filter generation unit 2401 sets the filter value of the frequency component to “1”.

Further, in the speech non-speech identification unit 106, when the power spectrum of the input speech signal is smaller than the multiplication result of the first threshold value for distinguishing speech and noise and the power spectrum of the input speech signal, that is, the expression When (21) is satisfied, the noise separation comb filter generation unit 2401 sets the value of the comb filter of the frequency component to “0”. Here, θ _nos is a threshold used for noise separation.

  The noise separation coefficient calculation unit 2402 multiplies the comb filter generated by the noise separation comb filter generation unit 2401 by an attenuation coefficient based on frequency characteristics, and sets the attenuation coefficient of the input signal for each frequency component. The attenuation coefficient of each frequency component is output to multiplication section 2403. Specifically, the noise separation coefficient calculation unit 2402 sets the noise separation coefficient sepn (k) = 1 when the value of the comb filter COMB_nos (k) is 0, that is, the blocking region.

If the value of the comb filter COMB_nos (k) is 1, that is, the pass region, the noise separation coefficient sepn (k) is calculated from the following equation (22).

Here, r _d (i) is a random function and is composed of uniformly distributed random numbers. Further, k is a variable for specifying a bin, and the range that k can take is 0 or more and less than half of the number of data subjected to FFT transform length, that is, fast Fourier transform.

  The multiplier 2403 multiplies the speech spectrum output from the frequency divider 104 by the noise separation coefficient output from the noise separation coefficient calculator 2402 in units of frequency components. Then, the spectrum obtained as a result of multiplication is output to noise frequency synthesis section 2404.

  The noise frequency synthesis unit 2404 synthesizes the spectrum of the frequency component unit output from the multiplication unit 2403 into a speech spectrum continuous in the frequency domain in units of a predetermined processing time, and outputs the synthesized speech spectrum to the IFFT unit 111. The IFFT unit 111 performs IFFT on the audio spectrum output from the noise frequency synthesis unit 2404 and outputs a signal converted into an audio signal.

  As described above, the speech processing apparatus according to the present embodiment discriminates the speech non-speech of the spectrum signal in units of frequency components, and attenuates the frequency characteristics based on the discrimination results in units of frequency components, thereby obtaining an accurate pitch. It is possible to create a comb filter that obtains information and extracts only noise components, and to extract noise characteristics. In addition, the noise component is not attenuated in the stop band of the comb filter, and a good noise separation characteristic can be obtained by reconstructing the noise component by multiplying the noise-based estimated value by a random number in the pass band of the comb filter. .

(Embodiment 14)
FIG. 25 is a block diagram of an example of the configuration of the speech processing apparatus according to the fourteenth embodiment. However, the same components as those in FIGS. 1 and 24 are denoted by the same reference numerals as those in FIGS. 1 and 24, and detailed description thereof is omitted.

  25 includes an SNR calculation unit 2501, an audio noise frame detection unit 2502, a noise comb filter reset unit 2503, and a noise separation comb filter generation unit 2504, and includes an audio component in an input audio signal. 1 and 24 is different from the speech processing apparatus shown in FIGS. 1 and 24 in that the frequency pass band of the noise separation comb filter for all frames is set as a stop band.

  The SNR calculator 2501 calculates the SNR of the audio signal from the first comb filter output from the audio spectrum output from the frequency divider 104, and outputs the calculation result to the audio noise frame detector 2502.

  The voice noise frame detection unit 2502 determines whether the input signal is a voice signal or a noise signal in units of frames from the SNR output from the SNR calculation unit 2501, and outputs the determination result to the noise comb filter reset unit 2503. Specifically, when the SNR is greater than a predetermined threshold, the audio noise frame detection unit 2502 determines that the input signal is an audio signal (audio frame), and the number of frames having an SNR equal to or less than the predetermined threshold is greater than or equal to a predetermined number. When it occurs continuously, the input signal is determined as a noise signal (noise frame).

  The noise comb filter reset unit 2503 sends a comb to the noise separation comb filter generation unit 2504 when the determination result in the audio noise frame detection unit 2502 is a determination result in which the audio component is not included in the frame of the input audio signal and only the noise component is included. An instruction to convert all frequency passbands of the filter into a stopband is output.

  Noise separation comb filter generation section 2504 generates a comb filter that emphasizes pitch harmonics based on the presence or absence of speech components in each frequency component, and outputs this comb filter to noise separation coefficient calculation section 2402.

Specifically, in the speech non-speech identification unit 106, when the power spectrum of the input speech signal is equal to or greater than the multiplication result of the first threshold value for determining speech and noise and the power spectrum of the input speech signal. That is, when Expression (20) is satisfied, the noise separation comb filter generation unit 2504 sets the filter value of the frequency component to “1”.

Further, in the speech non-speech identification unit 106, when the power spectrum of the input speech signal is smaller than the multiplication result of the first threshold value for distinguishing speech and noise and the power spectrum of the input speech signal, that is, the expression When (21) is satisfied, the noise separation comb filter generation unit 2504 sets the value of the comb filter of the frequency component to “0”. Here, θ _nos is a threshold used for noise separation.

  In addition, when the noise separation comb filter generation unit 2504 receives an instruction from the noise comb filter reset unit 2503 to convert all frequency passbands of the comb filter into a stopband, the noise separation comb filter generation unit 2504 blocks all the frequency passbands of the comb filter according to the instruction. Convert to area.

  As described above, according to the speech processing device of the present embodiment, when it is determined that the frame of the input speech signal does not include speech and is only a noise component, all frequency passbands of the comb filter are converted into stopbands. Thus, noise can be cut in the entire band in a signal section that does not include speech, and good noise separation characteristics can be obtained.

(Embodiment 15)
FIG. 26 is a block diagram of an example of the configuration of the speech processing apparatus according to the fifteenth embodiment. However, the same components as those in FIGS. 1 and 24 are denoted by the same reference numerals as those in FIGS. 1 and 24, and detailed description thereof is omitted. The speech processing apparatus of FIG. 26 includes an average value calculation unit 2601 and obtains the power average value of the speech spectrum in each frequency component or the power average value of the frame processed in the past and the frame processed. 1 and the speech processing apparatus of FIG.

The average value calculation unit 2601 takes the average value of the power of the speech spectrum output from the multiplication unit 2403 with the frequency components in the vicinity and the average value of the frames processed in the past, and uses the obtained average value for noise frequency synthesis. The data is output to the unit 2404. Specifically, the average value of the speech spectrum is calculated using the following formula (6).

  Here, k1 and k2 indicate frequency components, and k1 <k <k2. n1 indicates a number indicating a frame that has been processed in the past, and n indicates a number indicating a frame that has been processed.

  As described above, according to the speech processing apparatus according to the fifteenth embodiment of the present invention, the average power value of the speech spectrum in each frequency component or the average power value of the frame that has been processed in the past and the frame that has been processed is obtained. Thus, the influence of the sudden noise component is reduced.

(Embodiment 16)
FIG. 27 is a block diagram of an example of the configuration of the speech processing apparatus according to the sixteenth embodiment. However, the same components as those in FIG. 1 are denoted by the same reference numerals as those in FIG. The voice processing apparatus of FIG. 27 is an example in which voice enhancement and noise extraction are performed by combining the voice processing apparatus of FIG. 11 and the voice processing apparatus of FIG.

  In FIG. 27, the frequency dividing unit 104 divides the audio spectrum output from the FFT unit 103 into frequency components, and converts the audio spectrum for each frequency component into a noise-based estimation unit 1101, a first audio non-speech identification unit 1102, The data is output to the second voice / non-voice identification unit 1103, the voice pitch estimation unit 1104, the multiplication unit 2403, and the third voice / non-voice identification unit 2701.

  When the determination result that the speech component is included in the frame is output from the first speech non-speech identification unit 1102, the noise base estimation unit 1101 outputs the noise base estimated in the past to the first speech non-speech discrimination unit 1102. To do. Further, when the determination result that the speech component is included in the frame is output from the second speech non-speech identification unit 1103, the noise base estimation unit 1101 determines the noise base estimated in the past as the second speech non-speech discrimination unit 1103. Output to. Similarly, when the determination result that the speech component is included in the frame is output from the third speech non-speech discrimination unit 2701, the noise base estimation unit 1101 determines the noise base estimated in the past as the third speech non-speech discrimination unit. 2701 is output.

  In addition, the noise base estimation unit 1101 outputs a determination result that the frame does not include a voice component from the first voice non-voice identification unit 1102, the second voice non-speech discrimination unit 1103, or the third voice non-speech discrimination unit 2701. If it is, the moving average value representing the average amount of the short-time power spectrum and the spectrum change for each frequency component of the speech spectrum output from the frequency division unit 104 is calculated, and the moving average value and the power spectrum calculated in the past are calculated. Is used to calculate a new moving average value.

  When the difference between the speech spectrum signal output from the frequency division unit 104 and the noise base value output from the noise base estimation unit 1101 is equal to or greater than a predetermined first threshold, the first speech non-speech identification unit 1102 In other cases, it is determined that the sound part is a silent part including only noise. In the first voice / non-voice identification unit 1102, the first threshold value is used by the second voice / non-voice identification unit 1103 to be described later in order to generate a filter from which the first comb filter generation unit 1105 extracts as much voice pitch information as possible. Set to a value lower than the second threshold.

  Then, the first voice / non-voice identification unit 1102 outputs the determination result to the first comb filter generation unit 1105.

  When the difference between the speech spectrum signal output from the frequency division unit 104 and the noise base value output from the noise base estimation unit 1101 is equal to or greater than a predetermined second threshold, the second speech non-speech identification unit 1103 In other cases, it is determined that the sound part is a silent part including only noise. Then, the second voice / non-voice identification unit 1103 outputs the determination result to the second comb filter generation unit 1106.

  First comb filter generation section 1105 generates a first comb filter that emphasizes pitch harmonics based on the presence or absence of audio components in each frequency component, and outputs the first comb filter to comb filter correction section 1108.

  The voice pitch estimation unit 1104 estimates the voice pitch period from the voice spectrum output from the frequency division unit 104 and outputs the estimation result to the voice pitch restoration unit 1107. The voice pitch restoration unit 1107 corrects the second comb filter based on the estimation result output from the voice pitch estimation unit 1104 and outputs the result to the comb filter correction unit 1108.

  The comb filter correction unit 1108 corrects the first comb filter generated in the first comb filter generation unit 1105 using the pitch repair comb filter generated in the voice pitch repair unit 1107, and converts the corrected comb filter into a voice separation coefficient. The result is output to the calculation unit 1109.

  The speech separation coefficient calculation unit 1109 multiplies the comb filter modified by the comb filter modification unit 1108 by a separation coefficient based on the frequency characteristics, calculates a separation coefficient of the input signal for each frequency component, and outputs it to the multiplication unit 109. To do. The multiplication unit 109 multiplies the speech spectrum output from the frequency division unit 104 by the attenuation coefficient output from the speech separation coefficient calculation unit 1109 in units of frequency components. Then, the spectrum obtained as a result of multiplication is output to frequency synthesizer 110.

  The third speech non-speech identification unit 2701 includes a noise component when the difference between the speech spectrum signal output from the frequency division unit 104 and the noise base value output from the noise base estimation unit 1101 is equal to or greater than a predetermined threshold. It is determined as a voiced portion, and in other cases, it is determined as a silent portion of only noise that does not include a voice component. Then, the third speech non-speech identification unit 2701 outputs the determination result to the noise base estimation unit 1101 and the noise separation comb filter generation unit 2401.

  The noise separation comb filter generation unit 2401 generates a comb filter that emphasizes the voice pitch based on the presence or absence of a voice component in each frequency component, and outputs this comb filter to the noise separation coefficient calculation unit 2402. The noise separation coefficient calculation unit 2402 multiplies the comb filter generated by the noise separation comb filter generation unit 2401 by an attenuation coefficient based on frequency characteristics, and sets the attenuation coefficient of the input signal for each frequency component. The attenuation coefficient of each frequency component is output to multiplication section 2403.

  The multiplier 2403 multiplies the speech spectrum output from the frequency divider 104 by the noise separation coefficient output from the noise separation coefficient calculator 2402 in units of frequency components. Then, the spectrum obtained as a result of multiplication is output to noise frequency synthesis section 2404. The noise frequency synthesis unit 2404 synthesizes the spectrum of the frequency component unit output from the multiplication unit 2403 into a speech spectrum continuous in the frequency domain in a predetermined processing time unit, and outputs the synthesized speech spectrum to the IFFT unit 2702.

  IFFT unit 2702 performs IFFT on the audio spectrum output from noise frequency synthesis unit 2404 and outputs a signal converted into an audio signal.

  As described above, according to the speech processing apparatus of the present embodiment, it is possible to accurately discriminate the speech non-speech of the spectrum signal in units of frequency components and perform attenuation of the frequency characteristics based on the discrimination results in units of frequency components. Therefore, even if noise suppression is performed with a large attenuation, it is possible to perform speech enhancement with little speech distortion. At the same time, noise extraction can be performed.

  Note that the sound processing apparatus of the present invention is not limited to the example of the sound processing apparatus of the sixteenth embodiment, and each of the above embodiments can be applied in combination.

  In addition, although the speech enhancement and noise extraction according to any of the above embodiments has been described as a speech processing apparatus, this speech enhancement and noise extraction can also be realized by software. For example, a program for performing voice enhancement and noise extraction may be stored in advance in a ROM (Read Only Memory), and the program may be operated by a CPU (Central Processor Unit).

  In addition, the program for performing speech enhancement and noise extraction is stored in a computer-readable storage medium, the program stored in the storage medium is recorded in a RAM (Random Access Memory) of the computer, and the computer is executed according to the program. May be. Even in such a case, the same operation and effect as the above-described embodiment are exhibited.

  Further, the program for performing speech enhancement may be stored in a server, the program stored in the server may be transferred to the client, and the program may be executed on the client. Even in such a case, the same operation and effect as the above-described embodiment are exhibited.

  Further, the voice processing device according to any of the above embodiments can be mounted on a wireless communication device, a communication terminal, a base station device, or the like. As a result, voice during communication can be emphasized or extracted with noise.

The block diagram which shows the structure of the speech processing unit which concerns on Embodiment 1 of this invention. Flow chart showing the operation of the speech processing apparatus in the above embodiment The figure which shows the example of the comb filter produced with the speech processing unit in the said embodiment FIG. 3 is a block diagram showing an example of the configuration of a speech processing apparatus according to the second embodiment. FIG. 3 is a block diagram showing an example of the configuration of a speech processing apparatus according to a third embodiment. FIG. 4 is a block diagram showing an example of the configuration of a speech processing apparatus according to a fourth embodiment. FIG. 6 is a block diagram showing an example of the configuration of a speech processing apparatus according to a fifth embodiment. FIG. 9 is a block diagram showing an example of the configuration of a speech processing apparatus according to a sixth embodiment. The figure which shows the example of restoration | repair of the comb filter in the speech processing unit in the said embodiment FIG. 7 is a block diagram showing an example of the configuration of a speech processing apparatus according to a seventh embodiment. FIG. 9 is a block diagram showing an example of the configuration of a speech processing apparatus according to an eighth embodiment. Diagram showing an example of a comb filter Diagram showing an example of a comb filter Diagram showing an example of a comb filter Diagram showing an example of a comb filter Diagram showing an example of a comb filter Diagram showing an example of a comb filter Diagram showing an example of a comb filter Block showing an example of the configuration of a speech processing apparatus according to the ninth embodiment The figure which shows an example of the audio | voice noise judgment program of the audio | voice processing apparatus of this Embodiment FIG. 10 is a block diagram showing an example of the configuration of a speech processing apparatus according to the tenth embodiment. FIG. 12 is a block diagram showing an example of the configuration of a speech processing apparatus according to an eleventh embodiment. FIG. 12 is a block diagram showing an example of the configuration of a speech processing apparatus according to a twelfth embodiment. FIG. 13 is a block diagram showing an example of the configuration of a speech processing apparatus according to a thirteenth embodiment. FIG. 14 is a block diagram showing an example of the configuration of a speech processing apparatus according to a fourteenth embodiment. FIG. 15 is a block diagram showing an example of the configuration of a speech processing apparatus according to a fifteenth embodiment. FIG. 16 is a block diagram showing an example of the configuration of a speech processing apparatus according to a sixteenth embodiment. The figure which shows the example of the speech processing device using the conventional comb filter method Diagram showing attenuation characteristics of comb filter

Explanation of symbols

104 frequency division unit 105, 1101 noise base estimation unit 106 speech non-speech identification unit 107 comb filter generation unit 108 attenuation coefficient calculation unit 109, 2403 multiplication unit 110 frequency synthesis unit 401 noise section discrimination unit 402 noise base tracking unit 501 musical noise suppression Unit 502, 1108 Comb filter correction unit 601, 2012, 2601 Average value calculation unit 701 Section discrimination unit 702, 2301 Com filter reset unit 801 Audio pitch period estimation unit 802, 1107 Audio pitch restoration unit 1001 Automatic threshold adjustment unit 1102 First audio Non-speech identification unit 1103 Second speech non-speech discrimination unit 1104, 1903 Speech pitch estimation unit 1105, 2101 First comb filter generation unit 1106, 2103 Second comb filter generation unit 1109 Speech separation coefficient calculation unit 1 901, 2501 SNR calculation unit 1902, 2502 Voice noise frame detection unit 2102 First musical noise suppression unit 2104 Second musical noise suppression unit 2401 Noise separation comb filter generation unit 2402 Noise separation coefficient calculation unit 2404 Noise frequency synthesis unit 2503 Noise comb filter Reset unit 2504 Noise separation comb filter generation unit 2701 Third voice non-voice identification unit

Claims (27)

Frequency dividing means for dividing the audio spectrum of the input signal in predetermined frequency domain units;
Voice / non-speech identification means for identifying whether or not a voice component is included in a divided voice spectrum that is a voice spectrum frequency-divided by the frequency division means;
A comb filter generating means for generating a comb filter having a frequency region including a voice component as a pass band and a frequency region not including a voice component as a stop band based on the identification result of the voice / non-voice discrimination means;
Multiplying the comb filter by an attenuation coefficient based on a frequency characteristic sets an attenuation coefficient for each frequency domain, and multiplying each divided speech spectrum by the attenuation coefficient of the corresponding frequency domain results in the divided speech spectrum Noise suppression means for suppressing noise components of
And a frequency synthesizer for synthesizing the divided speech spectrum in which the noise component is suppressed into a speech spectrum continuous in the frequency domain.   The speech / non-speech identification unit may include a speech component included in the divided speech spectrum when a difference value between the power of the divided speech spectrum and a noise-based power that is a spectrum of the noise component is larger than a predetermined threshold. The speech processing apparatus according to claim 1, wherein when the difference value is equal to or less than the threshold value, it is determined that a speech component is not included in the divided speech spectrum.   Mean value calculating means for taking an average value of the power of the divided speech spectrum is provided, and the speech / non-speech identifying means is a difference between the average value of the power of the divided speech spectrum and a noise-based power that is a spectrum of a noise component. When the value is larger than a predetermined threshold value, it is determined that the divided sound spectrum includes a sound component, and when the difference value is equal to or less than the threshold value, it is determined that the divided sound spectrum does not include a sound component. The speech processing apparatus according to claim 2.   Noise base estimation means for updating a noise base in a frequency domain that does not include speech components based on a noise base average value estimated in the past and a weighted average value of the power of the divided speech spectrum is provided. The speech processing apparatus according to any one of claims 1 to 3.   5. The speech processing apparatus according to claim 1, wherein the noise suppression unit attenuates the divided speech spectrum in a stop band of the comb filter.   A pitch frequency estimating means for estimating a voice pitch frequency; and a pitch correcting means for correcting a pitch harmonic width of the comb filter based on the voice pitch frequency and the divided voice spectrum, and the noise suppressing means 6. The speech processing apparatus according to claim 1, wherein a noise component of the divided speech spectrum is suppressed using a comb filter in which the pitch harmonic width is corrected.   The threshold is increased when the number of frequency components in the pass band of the comb filter is greater than a predetermined number, and the threshold is increased when the number of frequency components in the pass band of the comb filter is less than or equal to the predetermined number. 4. The voice processing apparatus according to claim 2, further comprising a threshold adjustment unit for reducing the threshold value.   An interval discriminating unit that determines whether or not the audio spectrum of the input signal includes speech, and when the interval discriminating unit determines that the audio spectrum includes only a noise component that does not include an audio component, The voice processing apparatus according to claim 1, further comprising: a comb filter resetting unit.   9. The method according to claim 1, further comprising a musical noise suppression unit that sets all of the comb filters to a stop band when the number of frequency components in the pass band of the comb filter is equal to or less than a predetermined number. A voice processing apparatus according to claim 1. Frequency dividing means for dividing the audio spectrum of the input signal in predetermined frequency domain units;
First voice / non-voice identification means for identifying whether or not a voice component is included in a divided voice spectrum that is a voice spectrum frequency-divided by the frequency division means;
Based on the identification result of the first speech / non-speech discrimination means, a first comb filter is generated which uses a frequency region including a speech component as a pass region and a frequency region not including a speech component as a stop region. A comb filter generating means;
Second voice / non-voice identification means for identifying whether or not a voice component is included in the divided voice spectrum under different conditions from the first voice / non-voice identification means;
A second comb filter that generates a second comb filter having a frequency region including a voice component as a pass band and a frequency region not including a voice component as a stop band based on the identification result of the second voice / non-voice discrimination unit. A comb filter generating means;
Voice pitch estimation means for estimating a pitch frequency of an input voice signal from the divided voice spectrum;
Voice pitch repairing means for repairing the pitch harmonic width of the second comb filter based on the pitch frequency estimated by the voice pitch estimation means and generating a pitch repair comb filter;
Comb filter correcting means for correcting the first comb filter based on the pitch repair comb filter and generating a corrected comb filter;
Multiplying the modified comb filter by an attenuation coefficient based on a frequency characteristic sets an attenuation coefficient for each frequency domain, and multiplying each of the divided audio spectrums by the attenuation coefficient of the corresponding frequency domain Noise suppression means for suppressing noise components of the spectrum;
And a frequency synthesizer for synthesizing the divided speech spectrum in which the noise component is suppressed into a speech spectrum continuous in a frequency domain. The first speech / non-speech identification unit is configured to generate a speech component in the divided speech spectrum when a difference value between the power of the divided speech spectrum and a noise-based power that is a spectrum of a noise component is larger than a predetermined first threshold. It is determined that it is included, and when the difference value is equal to or less than the first threshold value, it is determined that a voice component is not included in the divided voice spectrum,
The second speech / non-speech identification means includes a speech component in the divided speech spectrum when a difference value between the power of the divided speech spectrum and the noise-based power is greater than a second threshold value that is greater than the first threshold value. The speech processing apparatus according to claim 10, wherein when the difference value is equal to or smaller than the second threshold value, it is determined that a speech component is not included in the divided speech spectrum.   The second speech / non-speech discrimination means includes an average value calculating means for taking an average value of the power of the divided speech spectrum, and the second speech / non-speech discrimination means is a noise-based power that is a spectrum of an average value of the divided speech spectrum and a noise component. When the difference value is greater than the second threshold, it is determined that the divided sound spectrum includes a sound component, and when the difference value is equal to or smaller than the second threshold, the sound component is included in the divided sound spectrum. The speech processing apparatus according to claim 11, wherein it is determined that the voice processing has not been performed.   SNR calculating means for calculating a voice-to-noise ratio from the power of the divided voice spectrum and the first or second comb filter, and a voice / noise frame for detecting a voice frame or a noise frame based on the voice-to-noise ratio. 13. A noise frame detection unit, wherein the voice pitch estimation unit estimates the pitch frequency when the voice / noise frame detection unit determines that the voice frame is detected. The speech processing apparatus according to any one of the above.   14. The speech processing apparatus according to claim 13, further comprising comb filter resetting means for setting all of the modified comb filters to a stop band when the speech / noise frame detecting means determines that the frame is a noise frame.   The comb filter correcting means sets a frequency component that overlaps with a frequency component in the pass band of the pitch repair comb filter among frequency components in the pass band of the first comb filter as a pass band of the corrected comb filter. The speech processing apparatus according to any one of claims 10 to 14, wherein a frequency region other than a frequency region is set as a stop region of the modified comb filter. When the number of frequency components in the pass band of the first comb filter is equal to or less than a predetermined number, first musical noise suppression means that makes all the first comb filter a stop band,
And a second musical noise suppression unit that uses the second comb filter as a stop band when the number of frequency components in the pass band of the second comb filter is a predetermined number or less. The voice processing device according to any one of claims 10 to 15. Frequency dividing means for dividing the audio spectrum of the input signal in predetermined frequency domain units;
Voice / non-speech identification means for identifying whether or not a voice component is included in a divided voice spectrum that is a voice spectrum frequency-divided by the frequency division means;
When the power of the divided speech spectrum is larger than the multiplication result of the noise-based power that is the spectrum of the noise component and a predetermined threshold, the frequency region is set as a passband, and the power of the divided speech spectrum is less than or equal to the multiplication result A comb filter generating means for generating a noise separation comb filter having the frequency region as a stop band;
The noise separation comb filter is multiplied by an attenuation coefficient based on frequency characteristics to set an attenuation coefficient for each frequency domain, and the divided speech spectrum is multiplied by the corresponding frequency domain attenuation coefficient to perform the division. A noise separation coefficient calculating means for extracting a noise component of the speech spectrum;
A speech processing apparatus comprising: frequency synthesis means for synthesizing the divided speech spectrum from which the noise component has been extracted into a speech spectrum continuous in the frequency domain.   18. The speech processing apparatus according to claim 17, wherein the noise separation coefficient calculation unit extracts the noise component by multiplying a noise-based estimated value by a random number in a pass band of the comb filter.   SNR calculating means for calculating a voice-to-noise ratio from the power of the divided voice spectrum and the first or second comb filter, and a voice / noise frame for detecting a voice frame or a noise frame based on the voice-to-noise ratio. A noise frame detection unit; and a comb filter reset unit that sets the noise separation comb filter to a stop band when the voice / noise frame detection unit determines that the voice spectrum includes only a noise component that does not include a voice component. The speech processing apparatus according to claim 17 or 18, further comprising:   19. The speech processing apparatus according to claim 17, further comprising spectrum averaging means for calculating a frequency average and a time average of a divided speech spectrum after speech processing using the noise separation comb filter.   A wireless communication apparatus comprising the voice processing apparatus according to any one of claims 1 to 20. A frequency division procedure for dividing the audio spectrum of the input signal in predetermined frequency domain units;
A voice / non-voice identification procedure for identifying whether or not a voice component is included in a divided voice spectrum that is a voice spectrum frequency-divided in the frequency division procedure;
A comb filter generation procedure for generating a comb filter based on the identification result of the speech / non-speech identification procedure, with a frequency region including a speech component as a passband and a frequency region including no speech component as a stopband;
Multiplying the comb filter by an attenuation coefficient based on a frequency characteristic sets an attenuation coefficient for each frequency domain, and multiplying each divided speech spectrum by the attenuation coefficient of the corresponding frequency domain results in the divided speech spectrum A noise suppression procedure for suppressing noise components of
And a frequency synthesis procedure for synthesizing the divided speech spectrum in which the noise component is suppressed into a speech spectrum continuous in the frequency domain. A frequency division procedure for dividing the audio spectrum of the input signal in predetermined frequency domain units;
A first voice / non-voice identification procedure for identifying whether or not a voice component is included in a divided voice spectrum that is a voice spectrum frequency-divided in the frequency division procedure;
A first comb filter that generates a comb filter based on the identification result of the first speech / non-speech identification procedure, having a frequency region including a speech component as a pass region and a frequency region not including a speech component as a stop region. Generation procedure,
A second voice / non-voice identification procedure for identifying whether or not a voice component is included in the divided voice spectrum under different conditions from the first voice / non-voice identification procedure;
A second comb filter is generated based on the identification result of the second speech / non-speech identification procedure, with a frequency region including a speech component as a pass region and a frequency region not including a speech component as a stop region. Com filter generation procedure,
A voice pitch estimation procedure for estimating a pitch frequency of an input voice signal from the divided voice spectrum;
A voice pitch restoration procedure for generating a pitch restoration comb filter by restoring a pitch harmonic width of the second comb filter based on the pitch frequency estimated in the voice pitch estimation procedure;
A comb filter modification procedure for modifying the first comb filter based on the pitch repair comb filter and generating a modified comb filter;
Multiplying the modified comb filter by an attenuation coefficient based on a frequency characteristic sets an attenuation coefficient for each frequency domain, and multiplying each of the divided audio spectrums by the attenuation coefficient of the corresponding frequency domain A noise suppression procedure to suppress the noise component of the spectrum;
And a frequency synthesis procedure for synthesizing the divided speech spectrum in which the noise component is suppressed into a speech spectrum continuous in the frequency domain.   23. A server that records the voice processing program according to claim 22, and transfers the voice processing program to a request source in response to a request.   24. A server that records the voice processing program according to claim 23 and transfers the voice processing program to a request source in response to a request. A frequency division step of dividing the audio spectrum of the input signal by a predetermined frequency domain unit;
A voice / non-voice identification step for identifying whether or not a voice component is included in the divided voice spectrum, which is the frequency-divided voice spectrum;
A pitch harmonic structure generation step for generating a pitch harmonic structure that emphasizes a frequency region including a speech component;
An attenuation coefficient setting step for setting an attenuation coefficient for each frequency region by multiplying the pitch harmonic structure by an attenuation coefficient based on a frequency characteristic;
A noise suppression step of suppressing a noise component of the divided speech spectrum by multiplying each of the divided speech spectra by the attenuation coefficient of the corresponding frequency domain;
And a frequency synthesis step of synthesizing the divided speech spectrum in which the noise component is suppressed into a speech spectrum continuous in a frequency domain. A frequency division step of dividing the audio spectrum of the input signal by a predetermined frequency domain unit;
A difference value calculating step of calculating a difference value between a power of a divided voice spectrum that is the frequency-divided voice spectrum and a noise-based power that is a spectrum of a noise component;
A first speech / non-speech identification step for identifying that a speech component is included in a divided speech spectrum in which the difference value is greater than a predetermined first threshold;
A first pitch harmonic structure generating step for generating a first pitch harmonic structure that emphasizes a frequency region identified as containing a voice component in the first voice / non-voice identification step;
A second speech / non-speech identification step for identifying that a speech component is included in a divided speech spectrum in which the difference value is greater than a second threshold greater than the first threshold;
A second pitch harmonic structure generating step for generating a second pitch harmonic structure that emphasizes the frequency region identified as including a voice component in the second voice / non-voice identification step;
A pitch frequency estimating step of estimating a pitch frequency of an input voice signal from the divided voice spectrum;
A third pitch harmonic structure generating step for generating a third pitch harmonic structure in which only peak information is extracted from the second pitch harmonic structure;
A fourth pitch harmonic structure generating step for generating a fourth pitch harmonic structure in which a peak is inserted in a portion corresponding to the estimated pitch frequency of the third pitch harmonic structure;
A fifth pitch harmonic structure generating step for generating a fifth pitch harmonic structure in which the peak width of the fourth pitch harmonic structure is widened according to the value of the pitch;
A sixth pitch harmonic structure generating step for generating a sixth pitch harmonic structure that emphasizes only a frequency region emphasized in both the first pitch harmonic structure and the fifth pitch harmonic structure;
An attenuation coefficient setting step for setting an attenuation coefficient for each frequency region by multiplying the sixth pitch harmonic structure by an attenuation coefficient based on a frequency characteristic;
A noise suppression step of suppressing a noise component of the divided speech spectrum by multiplying each of the divided speech spectra by the attenuation coefficient of the corresponding frequency domain;
And a frequency synthesis step of synthesizing the divided speech spectrum in which the noise component is suppressed into a speech spectrum continuous in a frequency domain.

Images